Transmission system anomaly detection method, system, medium and equipment

Abstract

The invention discloses a transmission system anomaly detection method and system, a medium and equipment, and the method comprises the steps: collecting vibration signals of a transmission system under different operation conditions through a vibration sensor; inputting the vibration signal into an auto-encoder for reconstruction; the working condition information is used as a label, and supervised comparative learning is introduced into an auto-encoder architecture; the distance between the test sample and the training sample center is used as an anomaly detection score to indicate whether the sample is abnormal or not; and adaptively obtaining probability density distribution of an anomaly detection score of a sample through kernel density estimation, and adopting a 5% quantile as a threshold value to realize anomaly detection of the transmission system.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment status, and particularly relates to a method, system, medium and device for detecting abnormalities in a transmission system. Background Art

[0002] The abnormal detection of the transmission system of high-end equipment is crucial for the safe operation of the whole machine. With the increasing complexity of mechanical equipment, its transmission system will face more complex application scenarios, resulting in faults such as wear, fatigue and abnormal displacement of the system. Most of the existing abnormal detection methods focus on the abnormal detection under stable working conditions, and for equipment working under different working conditions, their performance will decline. The change of working conditions will cause signal drift, resulting in false alarms or missed detections of the algorithm. In addition, the fixed detection threshold has insufficient robustness to the changing signal and fails to fully realize the detection performance of the algorithm.

[0003] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0004] The present invention provides a method, system, medium and device for detecting abnormalities in a transmission system, which learns the characteristics of the normal state of samples under different working conditions through a supervised contrastive learning method, and enhances the abnormal detection effect by using an adaptive threshold.

[0005] A method for detecting abnormalities in a transmission system includes:

[0006] Step a, using a vibration sensor to collect vibration signals of the transmission system under different operating conditions;

[0007] Step b, inputting the vibration signal into an autoencoder for reconstruction;

[0008] Step c, taking the working condition information as a label, introducing supervised contrastive learning into the autoencoder architecture to form test samples and training samples;

[0009] Step d, in the hidden layer feature space constrained by the supervised contrastive autoencoder, calculating the distance between each test sample and the center of all training samples as the abnormal score of the test sample to indicate whether the sample is abnormal;

[0010] Step e, adaptively obtaining the probability density distribution of the abnormal scores of the samples through kernel density estimation, and using the 5% quantile as the threshold to realize the abnormal detection of the transmission system.

[0011] In the described method for detecting abnormalities in a transmission system, in step a, the working conditions of the collected vibration signals include two types: loaded and unloaded.

[0012] In the described method for detecting abnormalities in a transmission system, in step b, the vibration signal is input into an autoencoder for training. The autoencoder consists of an encoder and a decoder . The encoder maps the vibration signal to a latent variable , and the decoder then generates reconstructed data from the latent variable . The autoencoder realizes data reconstruction by minimizing the reconstruction error between the vibration signal and the reconstructed data . The formula is as follows:

[0013] ,

[0014] ,

[0015] .

[0016] In the described method for detecting abnormalities in a transmission system, the vibration signal is first subjected to data augmentation. For a batch of N samples, each sample in is subjected to data augmentation twice to obtain two different views and , where , is the operating condition information of the normal sample. When the operating condition of the vibration signal is no-load, ; when the operating condition of the vibration signal is full-load, . Using the encoder as the backbone network, the feature representation of each view is extracted. The projection network maps this representation to the final vector , and then the final vector is normalized to the unit hypersphere, and then the supervised contrastive loss is calculated:

[0017] ,

[0018] where, is the index set of all positive examples different from in the multi-view batch, and is its cardinality.

[0019] In the described method for detecting abnormalities in a transmission system, the training set containing only normal data and the test set containing normal data and abnormal data are both input into the trained encoder to obtain their representations in the feature space. By calculating the test sample and its operating condition is the same One training sample Center The distance between them is used as the anomaly detection score :

[0020]

[0021] .

[0022] In the described method for detecting anomalies in a transmission system

[0023] Use the kernel density estimation non - parametric estimation method to estimate its probability density function from the anomaly detection scores. For each anomaly score , the kernel density value The estimation formula is as follows:

[0024] ,

[0025] Where is the number of samples, is the th sample, is the bandwidth, is the Gaussian kernel function,

[0026] Set the decision threshold according to the 5% quantile of the estimated probability distribution. Samples with a calculated probability density lower than the 5% quantile will be predicted as abnormal samples, and finally, anomaly detection of the transmission system under multiple working conditions is achieved.

[0027] A transmission system anomaly detection system includes

[0028] A vibration sensor that collects vibration signals of the transmission system under different operating conditions;

[0029] An auto - encoder that is connected to the vibration sensor to input the vibration signal into the auto - encoder for reconstruction;

[0030] A supervised contrast learning unit that is used to introduce supervised contrast learning into the auto - encoder architecture with the working condition information as the label;

[0031] An anomaly detection score generation unit that uses the distance between the test sample and the centroid of the training samples as the anomaly detection score;

[0032] A kernel function unit that adaptively obtains the probability density distribution of the anomaly detection scores of the samples through kernel density estimation and uses the 5% quantile as the threshold to achieve anomaly detection of the transmission system.

[0033] In the described system, the supervised contrast learning unit includes a processor.

[0034] A computer storage medium, the storage medium includes computer instructions, when it runs on a computer, it causes the computer to execute the described method.

[0035] An electronic device, the electronic device includes:

[0036] A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein,

[0037] When the processor executes the program, it implements the described method.

[0038] Compared with the prior art, the present invention has the following advantages: By using the working condition information as the label for supervised contrastive learning, the normal samples from the same working condition are more closely associated than the normal samples from different working conditions. Setting the threshold based on the probability distribution of kernel density estimation can more accurately reflect the data distribution and improve the accuracy of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] By reading the detailed description of the preferred specific embodiments below, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The accompanying drawings in the specification are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings. And throughout the drawings, the same reference numerals are used to represent the same components.

[0040] In the drawings:

[0041] Figure 1 is a flowchart of a method for detecting anomalies in a transmission system provided by an embodiment of the present disclosure;

[0042] Figure 2 is a graph of the probability density estimation result of the anomaly score provided by an embodiment of the present disclosure;

[0043] Figure 3 is a probability density distribution graph of the anomaly scores of normal samples and crack samples provided by an embodiment of the present disclosure;

[0044] Figure 4 is a probability density distribution graph of the anomaly scores of normal samples and tooth missing samples provided by an embodiment of the present disclosure.

[0045] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments. DETAILED DESCRIPTION OF THE INVENTION

[0046] Specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0047] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different terms to refer to the same component. The specification and claims do not use the difference in terms as a way to distinguish components, but use the difference in the functions of components as the criterion for distinction. As used throughout the specification and claims, the term "comprising" or "including" is an open-ended term and should be interpreted as "including but not limited to". The subsequent description in the specification is the preferred implementation manner for implementing the present invention, but the description is for the purpose of the general principles of the specification and is not used to limit the scope of the present invention. The scope of protection of the present invention shall be determined by the scope defined by the appended claims.

[0048] For the convenience of understanding the embodiments of the present invention, the following will further explain with specific embodiments as examples in conjunction with the accompanying drawings, and each accompanying drawing does not constitute a limitation on the embodiments of the present invention.

[0049] As Figures 1 to 4 shown, the method for detecting abnormal transmission system includes the following steps:

[0050] Step a, using a vibration sensor to collect vibration signals of the transmission system under different operating conditions;

[0051] Step b, inputting the vibration signal into an autoencoder for reconstruction;

[0052] Step c, using the condition information as a label, introducing supervised contrastive learning into the autoencoder architecture to form test samples and training samples;

[0053] Step d, in the hidden layer feature space constrained by the supervised contrastive autoencoder, calculating the distance between each test sample and the center of all training samples as the abnormal score of the test sample to indicate whether the sample is abnormal;

[0054] Step e, adaptively obtaining the probability density distribution of the abnormal scores of the samples through kernel density estimation, and using the 5% quantile as the threshold to achieve abnormal detection of the transmission system.

[0055] In the preferred implementation manner of the method for detecting abnormal transmission system, in step a, the operating conditions of the collected vibration signals include two types: with load and no load.

[0056] In the described method for detecting abnormalities in a transmission system, in step b, the vibration signal is input into an autoencoder for training. The autoencoder consists of an encoder and a decoder . The encoder maps the vibration signal to a latent variable . The decoder then generates reconstructed data from the latent variable . The autoencoder realizes data reconstruction by minimizing the reconstruction error between the vibration signal and the reconstructed data . The formula is as follows:

[0057] ,

[0058] ,

[0059] .

[0060] In a preferred embodiment of the described method for detecting abnormalities in a transmission system, in step c, first, data augmentation is performed on the vibration signal. For a batch of N samples, each sample in and is subjected to data augmentation twice to obtain two different views and , where is the operating condition information of the normal sample. When the operating condition of the vibration signal is no-load, ; when the operating condition of the vibration signal is full-load, . Using the encoder as the backbone network, the feature representation of each view is extracted. This representation is mapped to a final vector by the projection network. Then, the final vector

[0061] is normalized to the unit hypersphere, and then the supervised contrastive loss is calculated:

[0062] where is the index set of all positive examples different from in the multi-view batch, and is its cardinality.

[0063] In the described method for detecting abnormal transmission systems, in step d, the training set containing only normal data and the test set containing normal data and abnormal data are both input into the trained encoder to obtain their representations in the feature space. By calculating the distance between the feature of the test sample and the centroid of the training samples with the same operating conditions as the test sample, the distance is used as the anomaly detection score :

[0064]

[0065] .

[0066] In a preferred embodiment of the described method for detecting abnormal transmission systems, in step e,

[0067] the non-parametric estimation method of kernel density estimation is used to estimate its probability density function from the anomaly detection scores. For each anomaly score , the estimated formula for the kernel density value is as follows:

[0068] ,

[0069] where n is the number of samples, xi is the ith sample, h is the bandwidth, and K is the Gaussian kernel function.

[0070] According to the 5% quantile of the estimated probability distribution, a decision threshold is set. Samples with a calculated probability density lower than the 5% quantile will be predicted as abnormal samples, ultimately realizing the anomaly detection of the transmission system under multiple operating conditions.

[0071] A transmission system anomaly detection system includes,

[0072] a vibration sensor that collects vibration signals of the transmission system under different operating conditions;

[0073] an autoencoder that is connected to the vibration sensor to input the vibration signals into the autoencoder for reconstruction;

[0074] a supervised contrast learning unit that is used to introduce supervised contrast learning into the autoencoder architecture with the operating condition information as the label;

[0075] an anomaly detection score generation unit that is used to use the distance between the test sample and the centroid of the training samples as the anomaly detection score;

[0076] A kernel function unit, which is used to adaptively obtain the probability density distribution of the anomaly detection scores of samples through kernel density estimation, and uses the 5% quantile as the threshold to achieve anomaly detection of the drive system.

[0077] In a preferred embodiment of the system, the supervised contrast learning unit includes a processor.

[0078] A computer storage medium, the storage medium includes computer instructions, when it runs on a computer, it causes the computer to execute the described method.

[0079] An electronic device, the electronic device includes:

[0080] A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein,

[0081] When the processor executes the program, the described method is implemented.

[0082] In one embodiment, the drive system includes a rotating machine.

[0083] In one embodiment, as Figure 1 shown, the method for anomaly detection of a drive system based on contrast learning and adaptive threshold includes the following steps:

[0084] In the first step: Use a vibration sensor to collect vibration signals of the drive system under different operating conditions;

[0085] In the second step: Input the signal into an autoencoder for reconstruction;

[0086] In the third step: Use the condition information as a label and introduce supervised contrast learning into the autoencoder architecture;

[0087] In the fourth step: Use the distance between the test sample and the centroid of the training samples as the anomaly detection score;

[0088] In the fifth step: Adaptively obtain the probability density distribution of the sample anomaly scores through kernel density estimation, and use the 5% quantile as the threshold for anomaly detection.

[0089] In the first step of the method, the conditions of the collected signals include two types: on-load and no-load.

[0090] In the second step of the method, the signal is input into the autoencoder for training. The autoencoder consists of an encoder and a decoder The encoder maps the vibration signal to a latent variable The decoder then reconstructs the vibration signal from the latent variable Generate reconstructed data , the autoencoder minimizes the reconstruction error between the vibration signal and the reconstructed data to achieve data reconstruction. The formula is as follows:

[0091]

[0092]

[0093]

[0094] In the third step of the method, with the working condition information as the label, the supervised contrast learning loss is introduced into the autoencoder architecture. First, data augmentation is performed on the original data. For a batch of N samples, each sample in is augmented twice to obtain two different views and , where . In this study, is the working condition information of the normal sample. When the rotating manipulator is unloaded, ; when the rotating manipulator is fully loaded, . Using the encoder mentioned above as the backbone network, the feature representation of each view is extracted. This representation is mapped to the final vector by the projection network, and then is normalized to the unit hypersphere, and then the supervised contrast loss is calculated:

[0095]

[0096] where, is the index set of all positive examples different from in the multi-view batch, and is its cardinality.

[0097] In the fourth step of the method, the training set containing only normal data and the test set containing normal data and abnormal data are both input into the trained encoder to obtain their representations in the feature space. By calculating the distance between the feature of the test sample and the center of the training samples with the same working condition as it as the anomaly detection score :

[0098]

[0099] 。

[0100] In the fifth step of the method described above, the non-parametric estimation method of kernel density estimation is used to estimate the probability density function from the anomaly detection scores. For each anomaly score , the kernel density value is estimated by the following formula:

[0101] ,

[0102] where is the number of samples, is the -th sample, is the bandwidth, is the Gaussian kernel function.

[0103] The decision threshold is set according to the 5% quantile of the estimated probability distribution. Samples with a calculated probability density lower than the 5% quantile will be predicted as abnormal samples, finally realizing the anomaly detection of the drive system under multiple working conditions.

[0104] Figure 2 shows the probability density estimation results of the anomaly scores. The probability density distribution of the sample anomaly scores is adaptively obtained through kernel density estimation. Using the 5% quantile as the threshold for anomaly detection, the abnormal samples and normal samples are better distinguished, and both the false alarm rate and the missed alarm rate of the device are reduced.

[0105] Figure 3 shows the probability density distributions of the anomaly scores of normal samples and cracked samples. The features extracted by the above network make the probability density distributions of the anomaly scores of normal samples and cracked samples have less overlap, which is beneficial to anomaly detection.

[0106] Figure 4 shows the probability density distributions of the anomaly scores of normal samples and tooth-missing samples. The features extracted by the above network make the probability density distributions of the anomaly scores of normal samples and tooth-missing samples have less overlap, which is beneficial to anomaly detection.

[0107] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms under the inspiration of this specification and without departing from the scope protected by the claims of the present invention, and these all belong to the scope of protection of the present invention.

Claims

1. A method for detecting abnormality in a transmission system, characterized in that: The steps include: Step a, using a vibration sensor to collect vibration signals of the transmission system under different operating conditions; Step b, inputting the vibration signal into an autoencoder for reconstruction; Step c, using the working condition information as a label, introduces supervised contrastive learning into the autoencoder architecture and forms test samples and training samples; Step d, in the hidden feature space constrained by the supervised contrastive autoencoder, the distance between the center of each test sample and all training samples is calculated as the abnormality score of the test sample to indicate whether the sample is abnormal; In step e, the probability density distribution of the abnormality detection score of the sample is adaptively obtained through kernel density estimation, and the 5% quantile is used as the threshold to realize abnormality detection of the transmission system.

2. A transmission system abnormality detection method according to claim 1, characterized in that: Preferably, in step a, the working conditions of the collected vibration signal include loaded and unloaded conditions.

3. A transmission system abnormality detection method according to claim 1, characterized in that: In step b, the vibration signal is input into the autoencoder for training. The autoencoder consists of an encoder and decoder The encoder will vibrate the signal Mapping to latent variables , the decoder is then composed of hidden variables Generate reconstruction data , the autoencoder minimizes the vibration signal and reconstructing data The reconstruction error between To achieve data reconstruction, the formula is as follows: ， ， 。 4. A transmission system abnormality detection method according to claim 3, characterized in that: In step c, the vibration signal is firstly enhanced. For a batch of N samples, Each sample in Perform two data augmentations to get two different views and ,in , is the working condition information of the normal sample. When the working condition of the vibration signal is no-load, ; When the vibration signal is under full load, , using the encoder As the backbone network, each view is extracted The feature representation , which is mapped by the projection network to the final vector , and then the final vector Normalize to the unit hypersphere, and then calculate the supervised contrast loss: ， in, is different from The index set of all positive examples of is its cardinality.

5. A transmission system abnormality detection method according to claim 4, characterized in that: In step d, the training set containing only normal data and the test set containing normal data and abnormal data are input into the trained encoder to obtain their representations in the feature space. Same working condition as training samples Center The distance between them is used as the anomaly detection score : ， 。 6. A transmission system abnormality detection method according to claim 1, characterized in that: In step e, Using the kernel density estimation non-parametric estimation method, the probability density function is estimated from the anomaly detection score. For each anomaly score , kernel density value The estimation formula is as follows: ， in is the number of samples, For the samples, is the bandwidth, is the Gaussian kernel function, The decision threshold is set according to the 5% quantile of the estimated probability distribution. Samples with a calculated probability density lower than the 5% quantile will be predicted as abnormal samples, ultimately achieving abnormal detection of the transmission system under multiple working conditions.

7. A transmission system abnormality detection system, characterized in that: These include, Vibration sensor, which collects vibration signals of the transmission system under different operating conditions; An autoencoder connected to the vibration sensor to input the vibration signal into the autoencoder for reconstruction; A supervised contrastive learning unit, which is used to introduce supervised contrastive learning into the autoencoder architecture using working condition information as labels; an anomaly detection score generating unit, which is used to use the distance between the centroid of the test sample and the training sample as the anomaly detection score; The kernel function unit is used to adaptively obtain the probability density distribution of the abnormality detection score of the sample through kernel density estimation, and adopts the 5% quantile as the threshold to realize the abnormality detection of the transmission system.

8. The system according to claim 7, characterized in that The supervised contrastive learning unit includes a processor.

9. A computer storage medium, characterized in that The storage medium includes computer instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 6.

10. An electronic device, characterized in that: The electronic device comprises: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images